[SYCL][Doc] Device query for graphs support #178
Conversation
Although we have an emulation mode for when a SYCL backend doesn't support PI/UR command-buffers, we don't have good implementation coverage across it's usage on all possible backends. Providing a support query would allow us to limit the backends we support while an experimental extension to say CUDA and Level Zero, and maybe OpenCL. Rather than having users with FPGA and esimd_emulator trying to use the extension and crashing because we've not tested these platforms even with emulation mode. We could remove this query once the implementation is more mature and we can trust emulation mode support on all backends, but I don't think that's the case at the moment.
EwanC
added
Graph Implementation
|
This might be overcomplicating it, but would it be useful to instead of returning a |
Like that idea 👍 It's easy to provide the extra info, and I don't think it places any extra burden on users that don't care about emulated v native. |
EwanC
force-pushed
the
ewan/device_support_query
branch
2 times, most recently
from
2bba7e0 to
1ce0473
Compare
Pushed a commit with this change now. |
This provides extra information about how support is implemented.
EwanC
force-pushed
the
ewan/device_support_query
branch
from
1ce0473 to
bab9c2b
Compare
Implement the device info `graph_support` query defined by spec PR #178 This only reports that graphs are supported on Level Zero devices.
Implement the device info `graph_support` query defined by spec PR #178 This only reports that graphs are supported on Level Zero devices.
Implement the device info `graph_support` query defined by spec PR #178 This only reports that graphs are supported on Level Zero devices.
Although we have implemented an emulation mode for when a SYCL backend doesn't support PI/UR command-buffers, we don't have good implementation coverage across its usage on all possible backends.
Providing an extension support query would allow us to limit the backends we support while in experimental status. This could be to say CUDA and Level Zero, and maybe OpenCL, but we don't need to specify that here. Otherwise, users with backends we've not tested could try to use the extension and crash, even with emulation mode.
Longer term, I think we could remove this query once the implementation is more mature and we have faith in emulation mode support on all backends, but I don't think that's the case at the moment.